The next few years of AI

AI Futures · Dr Greg Baker

What becomes cheap.
What stays valuable.
What we do next.

An executive conversation · September 2026

Greg Baker

AI researcher and educator
Consulting CTO and software builder
Business owner, composer and habitual experimenter

I am interested in what changes when people can do things they could not afford to do before.

What decision brought you here?

Commitment

What are you about to invest in?

Assumption

What must remain true for it to work?

Horizon

When will you know whether you were right?

Keep the same decision in view throughout the session.

Expertise is becoming easier to reach

One person.
More capability.

Research and prepare a case
Organise evidence and arguments
Check the work; own the consequences

Greg’s experience illustrates access. It does not establish that AI caused a legal outcome.

Find the constraint

Where does work pile up?

Making every other step faster can simply produce a bigger queue.

What limits the customer’s result?

A slow decision?
A scarce expert?
A physical installation?
An unreliable handoff?

Three questions. Three kinds of evidence.

Capability

What can a system do under test?

Benchmarks

Use

What are people actually doing with it?

Product data

Impact

What changed for the business?

Measured outcomes

Longer tasks are becoming tractable

METR task horizons for selected models, showing 50 percent and 80 percent success estimates and uncertainty intervals on a logarithmic minutes axis.

Human-expert task duration. Selected models, not a complete frontier. Software, ML and cybersecurity tasks.

Reliability changes the decision

50%

Exploration

Retries may be cheap. Failure must be detectable.

80%

A different threshold

Still ask what failure costs and who checks.

Useful output − checking − rework − consequences

Software travels faster than machinery

Digital deployment

Copy and distribute
Connect to data
Change a workflow

Physical deployment

Install and integrate
Meet site constraints
Prove operating reliability

Forecasting becomes a routine input

What changes when a carefully researched forecast becomes cheap enough to ask for routinely?

The evidence

Several forecasting systems reached statistical parity with the superforecaster reference.

The business question

Which decision would improve if you updated its probability every month?

July 2026 ForecastBench comparison. Tool-using systems; overlapping uncertainty; extrapolated 2024 human reference.

Prepare for faster, slower and uneven change

Faster

Capability improves and customers adopt quickly.

Can you respond?

Slower

Reliability, economics or implementation disappoint.

Can you afford to wait?

Uneven

One part of the value chain changes first.

Who captures the benefit?

AI is entering the process of building AI

Research ideas→Code & experiments→Evaluation↺

A possible accelerator, with constraints.

Compute and energy
Useful experiments and reliable evaluation
Deployment and human institutions

What forecast would change your decision?

Probability of the event: 50%

Cost of acting now ($)Loss avoided if the event occurs ($)

Illustrative expected-loss model. Assumes acting prevents the stated loss, with no other benefits or costs.

Try SignalStorm →

When cognition gets cheap, what becomes scarce?

Output needs more than intelligence.

Access to customers
Physical capacity and location
Permissions and responsibility
Context, trust and coordination

Candidate constraints to test, not guaranteed winners.

Historical portrait of Karl Marx, retained from Pilot-01 for the discussion of production inputs

The parts that do not speed up

Baumol’s cost disease

A live string quartet still needs musicians and time.

As productivity changes elsewhere, the relative cost of labour-intensive work can rise.

What still takes time in your business?

Violinist performing, the original Pilot-01 artwork

You have 100,000 free workers

So do your competitors.

They are inexperienced. They need instructions. Somebody has to check.

What changes when labour is plentiful?

Possible

What previously impossible work can you attempt?

Affordable

What expensive service becomes ordinary?

Personal

What can be tailored to each customer?

Value

What no longer earns a premium?

And who manages and checks all of that work?

Scarcity is only the first question

What remains scarce?
Who can capture the value?
What price are you paying for it?

Substitution, competition, financing and regulation can change the answer.

Test the business, not the job title

Software

Production gets cheaper.

Test distribution, integration and switching.

Physical operations

Reasoning gets cheaper.

Test installation, reliability and capacity.

A place to visit

Promotion gets cheaper.

Test discovery, experience and substitutes.

Stress-test the investment

Question	Faster	Slower	Uneven
Customer demand	What changes?	What persists?	Who moves first?
Margin & constraint	Where is value?	What still costs?	Who captures it?
Capital at risk	How much must you commit before learning?
Next evidence	What result would justify the next commitment?

Open the decision workbook →

What would make you change your mind?

Observe

One leading indicator.

Commit

One staged investment.

Reconsider

One result that changes the decision.

A plan needs a way to notice that its assumptions have failed.

Build what differentiates you

Buy what has become routine. Keep learning where customers care.

Different stages need different management

Genesis

Explore whether it works.

Custom

⚒

Learn with specific users.

Product

Compare and improve.

Utility

Control cost and reliability.

Lower cost here. Add customer value there.

Two explicit arrows on one map: standardise a supporting component to the right; improve a customer-facing offer towards greater differentiation.

Both moves can belong to the same business strategy.

Map one customer outcome

1 Name the customer’s need.

2 List the capabilities it depends on.

3 Place each from novel to routine.

4 Find the constraint and choose a move.

Where does your advantage actually come from?

Match the supplier to the bet

Evidence

Show it working on our task.

Economics

Include checking, integration and ongoing cost.

Continuity

Can we keep operating and take our data with us?

Buy an outcome you can verify.

Exceptional capability can become ordinary

SpecialistExpert setup

FrontierPaid access

Widely availableCheap or open

LocalWhere hardware permits

A possible diffusion path. Timing varies by task, economics, licence and hardware.

Your customers and staff get these tools too

More people can prepare a serious challenge.

Check a decision
Compare alternatives
Organise evidence
Ask better questions

How well can your organisation explain itself?

AI can change who contributes expertise

791 professionals at P&G

Solution quality

AI-assisted individuals matched unaided pairs in the experiment.

Selection

Choosing the best idea remained a separate difficulty.

One-day innovation tasks in one company. Published June 2026. Not a headcount rule.

Which handoff still earns its place?

Customer need→Draft / design→Review→Deliver

Keep

Independent challenge, authority and responsibility.

Test

Queues created only by lack of access to expertise.

Try RoleShock →

Will staff tell you what they have automated?

If saving time threatens someone’s job, hiding the saving can be a sensible response.

Ask

What did you change, and what still needs checking?

Offer

Time to improve the process and learn the next task.

Make a promise you can keep.

One strategic choice. One reversible test.

The next five years

Which advantage will you strengthen?

The next thirty days

Which assumption will you test?

Write the decision →

State the result that changes the decision

Owner & review date	Who decides, and when?
Baseline	What happens now?
Whole outcome	Time, quality, checking, rework and customer effect.
Decision rule	Expand, revise or stop at a stated result.

What will you do differently on Monday?

A decision.
A test.
A reason to change your mind.

AI Futures · Dr Greg Baker
ai-futures.industrial-linguistics.com